A Comprehensive Machine Learning Approach for EOR Screening in Sandstone and Carbonate Reservoirs

Abstract

Abstract Enhanced oil recovery (EOR) screening is a process that allows the selection of the best EOR technique for boosting oil recovery from specific reservoirs based on several criteria. With the recent advances in EOR techniques, conventional screening guidelines, including classical look-up tables, might lead to considerable financial and technical risk and uncertainty. This study discusses this problem and its drawbacks and further employs machine learning (ML) algorithms to develop comprehensive EOR screening guidelines for sandstone and carbonate reservoirs. This work applied ML algorithms to relate key reservoir parameters to the various EOR methods, including chemical, solvent, and thermal EOR categories, in sandstones and carbonates. These key reservoir parameters include oil viscosity, oil gravity, temperature, depth, lithology, porosity, permeability, thickness, pressure, oil saturation, and salinity. Some of these parameters are not found as criteria in similar studies. A large worldwide database of EOR projects was collected based on available literature and recent EOR surveys and used to train both supervised and unsupervised ML models to assess the best EOR strategy. 70% of the dataset was used for training and validation while the remaining 30% was kept for blind testing, ensuring the model's generalization capacity. The results showed that the Random Forest algorithm outperformed other classification ML approaches, including Naïve Bayes, tree-based models, and neural networks, in predicting the best EOR method with almost 90% accuracy, while Naïve Bayes achieved the lowest accuracy of 75% among the tested algorithms. Moreover, adding features like salinity, pressure, and thickness improves the robustness of the EOR screening model. While it adds a level of complexity and variability, it does not affect the performance of the prediction. Furthermore, the unsupervised clustering approach improved the Naïve Bayes algorithm's performance, but not the others. Additionally, the work showed that the highly imbalanced distribution of target classes results in considerable shortcomings if not addressed. Finally, the model and its limitations were validated through a detailed sensitivity analysis of feature-class type interactions and EOR domain knowledge. This study is one of the very few that employ ML for EOR screening. This paper addresses shortcomings of previous studies using a comprehensive dataset, including overlooked important EOR-related features, and implementing modern ML and deep-learning algorithms with multiple performance metrics to confirm their efficient utilization, yielding more accurate predictions in a broad range of reservoir properties. This study addresses the imbalanced dataset problem, reducing its uncertainty while predicting specific EOR methods, resulting in a better-generalized model. Based on the proposed approach in this study, a more reliable and quicker EOR screening decision can be made that de-risks and reduces the uncertainty in related field-scale implementations while assessing each feature's impact on the specific EOR methods.